Multilingual conversational dataset used to fine-tune HAAIE, a local AI companion designed to help with social anxiety and loneliness.
The goal is a companion that sounds human: warm when things are heavy, funny when they're not, honest when a user describes something risky, and never clinical.
| File | Language | Examples |
|---|---|---|
dataset_en.jsonl |
English (source) | 1 222 |
dataset_fra.jsonl |
French | 1 223 |
dataset_spa.jsonl |
Spanish | 1 223 |
dataset_por.jsonl |
Portuguese | 1 223 |
dataset_deu.jsonl |
German | 1 223 |
dataset_jpn.jsonl |
Japanese | 1 223 |
The English set is the original; the other languages were machine-translated with NLLB-200 (1.3B distilled), so expect some awkward phrasing.
Alpaca-style JSONL, one example per line:
{"instruction": "Can you help me?", "input": "", "output": "Always. What's going on?"}instruction— what the user saysinput— unused (empty)output— the companion's reply
The data mixes emotional support (loneliness, self-hatred, feeling invisible), casual small talk, and light absurd humor, so the model doesn't become a sad-only bot.
| Script | What it does |
|---|---|
transform.py |
Translates the English dataset into other languages locally with NLLB-200 |
get_feeling.py |
Detects emotions in a message (28 GoEmotions labels) with a multilingual classifier |
prompter.py |
Builds a short system prompt adapted to the detected emotion |
add_prompt_to_dataset.py |
Adds that emotional system prompt to each example of a .jsonl file |
python add_prompt_to_dataset.py dataset_en.jsonl
# → dataset_en_enriched.jsonl- This dataset is experimental and was not reviewed by mental-health professionals.
- A model trained on it is not a substitute for therapy or crisis support. If you are in danger, contact your local emergency number or a crisis line.
- Some replies use informal or rough language on purpose; filter them if your use case needs a neutral tone.